Costly and time-intensive trial-and-error sampling procedures plague traditional denim fabric production. This study develops and compares statistical and machine learning models to predict key properties of denim fabrics, aiming to reduce reliance on physical sampling in the textile manufacturing process. Models based on the two-level fractional factorial method using DOE, Random Forest (RF), and multilayer neural networks are constructed and rigorously validated using a real-world dataset comprising 1373 entries of denim fabric. Fabric properties, including before- and after-wash weights and thread counts, were accurately predicted. The neural network model demonstrated exceptional performance in predicting BW-EPI, achieving an R 2 of 0.979. In contrast, the DOE (multi-way) and RF models achieved strong prediction accuracy for AW-PPI, scoring R 2 of 0.971 and 0.969, respectively, accompanied by low MAE values (~0.77–0.90) and MAPE values (~1.34%–1.59%). The study further demonstrates practical user interfaces for both forward and reverse prediction, facilitating rapid, data-driven decision-making in denim production. Despite facing challenges in predicting fabric shrinkage due to dataset constraints, both statistical and machine learning models have proven highly effective for predicting fabric parameters, enabling data-driven decisions that improve resource utilization, enhance sustainability, minimize sampling trials, and accelerate development cycles.
Islam et al. (Wed,) studied this question.